Skip to main content
October 20, 2026
Available to every organization. The Agent Sandbox is free and is now the way every agent is created and edited.
Setting up an AI agent used to mean writing a long prompt, saving it, running it, and only then finding out whether it did what you meant. The work of describing the agent and the work of checking it were in different places, so getting an agent right took more round trips than it should have. The Agent Sandbox brings both into one workspace. You describe what you want the agent to do, and a chat drafts it for you. You test it on a handful of your own products and see exactly what it generates. When something is off, you say so in plain language, and the chat rewrites the prompt. Everything you need to shape the agent, its prompt, its checks, its options and dependencies, sits in tabs next to the results.

​
Describe it, do not draft it

When you create an agent, you can tell the chatbot your intent and let it draft the whole thing, or set it up by hand if you already know exactly what you want. From there the chat is how you steer. Ask for shorter descriptions, a different opening, or a rule about competitor brands, and it rewrites the prompt for you. You can point at a single product in the results with an @ mention when one output is off and the rest look fine.

​
See the real output first

The Test tab runs the agent on real products from your catalog, not synthetic samples. Results stream in one at a time, each with the value the agent generated, the reasoning behind it, and the source data it used. You can watch the agent think as it works, and open a detail view to see the exact prompt the model received. What you see here is what enrichment will produce at scale.

​
Every change is a version

Each rewrite is saved as a version in the Changes tab. Compare a wording against an earlier one, and click Go to this version to return to the one that worked better. When the output is where you want it, you create the agent, and it is ready to link to attributes and run across your catalog.

​
Why it matters

An agent is only as good as the loop you use to build it. By putting the description, the test run, and the refinement in one place, the sandbox makes that loop short: change the prompt in a sentence, re-run on the same products, see the difference. You spend your time judging output instead of guessing at phrasing.

Read the step-by-step guide

The full reference walks through the launch screen, the configuration tabs, testing on real products, refining through chat, and the Changes history.